Which one should I pick for a new project?

I need more context to give you a meaningful recommendation. These are two fundamentally different projects solving different problems:

deepseek-ai/flashmla is a library of optimized CUDA attention kernels for large language model inference and training. It provides:

• Sparse and dense attention implementations (Multi-head Latent Attention)
• Kernels for prefill and decoding stages
• FP8 KV cache support for efficient decoding
• Performance up to 660 TFlops on H800 GPUs [README excerpt]

microsoft/bitnet is a quantized LLM inference framework built on llama.cpp. It provides:

• Low-bit quantization (ternary, I2_S embeddings)
• CPU/GPU inference optimizations
• Model conversion and benchmarking tools

Choose FlashMLA if:

• You're building a high-performance inference or training system and need optimized attention kernels
• You're working with NVIDIA H800/H100 GPUs
• You need dense or sparse attention with FP8 precision

Choose BitNet if:

• You need to run LLMs on CPU or with extreme quantization
• You want a complete inference framework (not just kernels)
• You're optimizing for model size and memory footprint over raw speed

Project health: FlashMLA shows strong signals (RepoPilot grade A, 0 import cycles, well-maintained). BitNet has more limited visibility in the provided data but appears to be a working inference framework.

What's your use case—optimizing inference speed, reducing model size, or something else?

FOLLOW_UPS:

• Are you optimizing for GPU inference speed or CPU efficiency?
• Do you need training support or inference-only?
• What hardware do you target (NVIDIA, CPU, or both)?